|
Bugzilla – Full Text Bug Listing |
| Summary: | RAM usage reduction through reduction of retransmission events | ||
|---|---|---|---|
| Product: | ns-3 | Reporter: | natale.patriciello |
| Component: | tcp | Assignee: | natale.patriciello |
| Status: | RESOLVED MOVED | ||
| Severity: | enhancement | CC: | ns-bugs, pdbarnes, tomh, tommaso.pecorella |
| Priority: | P5 | ||
| Version: | pre-release | ||
| Hardware: | PC | ||
| OS: | Linux | ||
| Attachments: | Event saving patch | ||
I assume that you did check that the code is faster and that it does produce equivalent traces. If te assumption holds tue, I have no objections to the changes. I'd only add a small note in a moment explaining why the "strange" approach was chosen. As a side note... if (Simulator::Now () + m_rto > Simulator::Now() + Simulator::GetDelayLeft (m_retxEvent)) change to if (m_rto > Simulator::GetDelayLeft (m_retxEvent)) and please initialize m_realRTO in the constructors (it silences the warning). |
Created attachment 2188 [details] Event saving patch Currently, for each ACK receive the implementation cancel the old retransmission events and insert a new events, since the retransmit timer expiration time is moved forward in the future. One approach to reduce RAM and execution time is to avoid such cancel / insert with a "Replace" approach. It could be handled in core, but I don't know how much is feasible. What I propose is a TCP-only patch, which works in this way: * New ack arrives * -> Calculate the time at which it should expire (m_realRTO=now+m_rto) -> if m_retxEvent < m_realRTO, don't cancel/set anything, set a flag to false -> if m_retxEvent > m_realRTO, cancel the old event, and set a new RTO event, flag to true ** Rto expires ** -> Check the flag; if it is set, it means that the triggered event is not a real RTO. Set another retx event, at time m_realRTO - now. -> If it is not set, RTO is a real rto and it is expired; do the retransmission.